Together AI Expands Fine-Tuning Service with New Features
Overview
Together AI has announced significant feature expansions for “Together Fine-Tuning," its fine-tuning service designed for open-weight models. The updates include the addition of new models, the introduction of live experiment tracking, enhanced control features for training and data processing, and price reductions for select models.
Announcement Details
The update introduces several new features and improvements aimed at enhancing overall workflow efficiency and flexibility.
- New Model Additions: A wide range of open-weight models has been added to the supported list, including GLM-5.3, GLM-5.2, GLM-5.1, DeepSeek-V4-Flash-0731, DeepSeek-V4-Flash, Kimi K2.7-Code, Kimi K2.6, Qwen 3.8-27B, Qwen 3.6-35B-A3B, Qwen 3.6-27B, the Qwen 3.5 series, and the Gemma 4 series.
- Live Experiment Tracking: Metrics (such as loss, gradient norm, and learning rate) are now recorded at every training and evaluation step, allowing users to view and compare them in real time via the Together API, CLI, and UI dashboard.
- Introduction of Expert LoRA: LoRA adapters can now be applied directly to the expert layers of Mixture of Experts (MoE) models, improving the acquisition and retention of new knowledge compared to applying them only to attention layers.
- Early Stopping: A mechanism has been introduced to automatically stop training when the validation loss plateaus, retain the optimal checkpoint, and automatically refund unused training steps.
- Price Revision: Accompanied by training cost optimizations, prices have been reduced for many supported models. Reductions range from 30% up to 70% for models such as the gpt-oss series.
- Improved Data Processing Transparency: Features have been added to preview and validate tokenized data prior to training, apply sample weighting per JSONL file, and automatically validate files upon upload completion.
Impact on Local LLM Users
For engineers who fine-tune and operate open-weight models for local environments or custom tasks, this update directly contributes to streamlining validation cycles and reducing costs.
- Easy Integration of Latest Models: Users can execute fine-tuning using the latest open-weight models such as GLM, Qwen, DeepSeek, and Gemma directly on the platform without having to configure complex architecture-specific settings themselves.
- Lower Training Costs: Pricing for LoRA adapter training per million tokens has been reduced overall. For example, SFT for Qwen/Qwen3.5-9B drops from 0.48 to 0.34, and SFT for openai/gpt-oss-120b drops from 5.00 to 2.50, lowering the financial barrier for experimentation (price units conform to the source).
- Improved Debugging and Experiment Management: Upfront data validation, tokenization previews, mid-training metric monitoring, and the elimination of wasted costs via early stopping help shorten the time it takes to detect failures.